Why do AI procurement pilots fail?

AI pilots in procurement usually fail after the demo, not during it. The agent reads purchase requests, matches invoices and drafts orders well enough on a sample. Then it meets the real supplier master, the approval rules nobody wrote down, and a budget it can commit without limit. The model was rarely the problem. The work around it was.

In one sentence: a procurement pilot proves the agent can read the paperwork, and then fails on the parts of procurement that were never paperwork.

/01The general pattern, with numbers

Procurement is not special here. MIT's 2025 study of enterprise generative AI found about 95% of pilots produced no measurable profit and loss impact¹². Of the organisations that evaluated enterprise tools, it reported that 60% looked, 20% reached a pilot and 5% reached production³. The cause it named was not model quality but integration: tools that "don't learn from or adapt to workflows"¹.

Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027. It names three reasons: escalating costs, unclear business value and inadequate risk controls⁴. All three show up in procurement first, because procurement is where an agent's mistakes turn into money.

/02Five reasons procurement pilots stall

1. /01

   The supplier master is worse than anyone admits. Duplicate vendors, old bank details, three spellings of the same company. A person works around it from memory. An agent takes it literally, and the pilot sample was usually cleaned by hand before the demo.

2. /02

   The approval matrix lives in people's heads. Who can sign a PO over 10,000? What happens when the budget holder is on leave? In most firms the honest answer is "it depends, ask Priya". An agent cannot ask Priya, so either it guesses or every case comes back to a person and the pilot saves nothing.

3. /03

   Nobody capped commitments. Teams meter the model's token bill and forget the money the agent can commit: orders raised, discounts accepted, invoices released. That second number does not appear on any AI invoice, so it is the one that ends the pilot when something goes wrong.

4. /04

   The agent prepares and approves the same thing. A pilot is easiest to build when one agent raises the order, checks it and posts it. That quietly breaks segregation of duties, and internal audit stops the rollout as soon as they see it.

5. /05

   No one owns it after go-live. The pilot had a project team. Production has a queue. When supplier formats change or a new category arrives, quality drifts slowly, people take the work back, and six months later the licence renews for a tool nobody uses.

/03Build or buy changes the odds

MIT reported that buying from specialist vendors and building partnerships succeeded about 67% of the time, while internal builds succeeded about one third as often¹. It also found the largest returns in back-office work, not in sales and marketing, where more than half of budgets went¹.

That fits procurement. The model is a commodity. What takes time is the connectors to your ERP, the approval rules and the record auditors will ask for. A vendor that has done that before has already paid for the mistakes. If you buy, ask the questions that separate a system from a demo before you sign. Read the model vendor's terms like any supplier's: frontier model vendors as suppliers.

/04What to fix before the next pilot

• Write the approval matrix down as rules a system can check: amount bands, categories, named approvers and a stand-in for each. If you cannot write it, the agent cannot follow it.
• Cap commitments in money, per order and per day, with anything over the cap held for a named person. How to set AI agent spend limits covers the numbers.
• Split the duties. The agent may prepare. Someone else approves, and the record shows both names.
• Pilot on the messy cases, not a cleaned sample. Take a week of real requests, exceptions included, and measure how many the agent has to hand back.
• Name the owner for month seven before month one starts. The person who answers for the queue should own the agent working it.

None of this is unusual. It is the ordinary discipline of putting anything into production, applied to software that can give a different answer to the same purchase request. For the wider pattern beyond procurement, read why AI projects fail.

/05Gatehouse fit

Gatehouse, Surehand's control plane, is built around the three procurement traps above. The approval matrix becomes signed rules the agent is checked against before each action. Spend limits are set in money. Anything over a limit, or any changed bank detail, holds for a named approver, and every run ends in a sealed record. The composite invoice matching deployment shows the split in an accounts payable run.

/06At a glance